Clarification on transition variables for fixed effects models

[Laura Tan]

I have a question regarding preparation before performing fixed effects models. I conducted fixed effects analysis in Stata 18 using the code xtreg varlist, fe. Before this, I set up the data with xtset id wave. My question is whether I need to specify the transition of my independent variable. For example, I am examining the association between marital transition and changes in health. Should I generate a transition variable like this:

gen marital_ch = (10*maritalstatus) + L.maritalstatus
xtreg health i.marital_ch, fe

From this, I get the following transitions in marital_ch:

• Married to married
• Married to unmarried

Or should I simply include the marital status variable directly in the regression model?

xtreg health i.maritalstatus, fe

The marital status variable is binary. I have already included the fixed effects and set up xtset. Stata can automatically capture the transitions for me.

[Carlo Lazzaro]

Laura:
I would go -xtreg health i.maritalstatus, fe-, adding other predictors, if available, and considering -vce(cluster panelid)- standard errors if you have at least 30 panels.
As -i.maritalstatus- is time-varying, yoiu should not have problems in determining its coefficient.

[George Ford]

If you think divorce may be a factor, then you can compute the difference in martial status between periods. It would be calculated as the difference in current and lagged value.

So if married is a dummy = 1 for married and 0 for unmarried, you'd have

g divorce_recent = (married==0)*(L.married==1)

[Clyde Schechter]

I wouldn't be comfortable with simply using a dichotomous variable for change in marital status (nor, equivalently, just using the marital status variable itself in the model). Such a model implicitly assumes that the effect on the outcome of getting married is the same magnitude with opposite sign as the effect of getting unmarried. There are probably few outcomes for which this is true. On top of that, the change from married to unmarried might have different effects depending on whether it occurs by divorce or by widowing. Assuming your data includes information on that type of transition, I would treat them as different. In short, data permitting, I would create and use a variable with four levels: no change, got married, got divorced, and got widowed.